TLDR: A new research paper introduces a ‘Region-Aware Reconstruction Strategy’ for pre-training fMRI foundation models. By selectively masking specific anatomical brain regions using the AAL3 atlas, instead of random masking, the model learns more discriminative representations. This approach, evaluated on the ADHD-200 dataset, improved classification accuracy by 4.23% in distinguishing ADHD from healthy controls, with the limbic regions and cerebellum showing significant contributions. The findings suggest that incorporating anatomical structure into model pre-training enhances both performance and interpretability in neuroimaging.
The field of neuroimaging is witnessing a transformative era with the emergence of foundation models, which are designed to learn generalizable representations from vast and diverse brain imaging datasets. These models hold immense promise for understanding complex brain functions and disorders like Attention-Deficit/Hyperactivity Disorder (ADHD).
Functional Magnetic Resonance Imaging (fMRI) is a powerful tool for measuring brain activity, but its high dimensionality and variability across individuals and acquisition methods pose significant challenges for analysis. Traditional approaches to training these models often rely on self-supervised learning, particularly reconstruction-based methods where parts of the input data are masked, and the model learns to reconstruct them. While effective, many existing fMRI models apply masking after averaging voxel signals into region-of-interest (ROI) level time series, which can lead to a loss of fine-grained spatial information.
A recent study introduces an innovative approach to pre-training fMRI foundation models, moving beyond conventional random masking strategies. Researchers from St. Jude Children’s Research Hospital and The University of Memphis have developed a “Region-Aware Reconstruction Strategy” that directly applies ROI-guided masking to full 4D fMRI volumes. This method selectively masks semantically coherent brain regions using the Automated Anatomical Labelling Atlas (AAL3), ensuring that the model focuses on reconstructing meaningful neural signals within anatomically defined areas.
The core idea is to integrate anatomical regions directly into the masking process. Instead of randomly masking arbitrary parts of the brain, this strategy targets specific regions like the frontal, temporal, parietal, and occipital lobes, as well as the cerebellum, limbic regions, and subcortical structures. This preserves the detailed voxel-level spatial information, encouraging the model to learn localized and functionally relevant features.
To evaluate this new strategy, the researchers utilized the ADHD-200 dataset, which includes resting-state fMRI scans from 973 subjects, comprising both healthy controls and individuals diagnosed with ADHD. They adopted NeuroSTORM, a state-of-the-art foundation model pre-trained on over 28 million fMRI frames, as their foundational architecture. The model was trained to reconstruct the masked segments of the input sequence without relying on supervision, followed by a fine-tuning phase for prediction tasks.
The results demonstrated a significant improvement: the region-aware masking strategy achieved a 4.23% increase in classification accuracy for distinguishing healthy controls from individuals with ADHD, compared to conventional random masking. Specifically, masking the cerebellum and limbic regions yielded the most substantial improvements in both classification accuracy and AUCROC (Area Under the Receiver Operating Characteristic curve). This indicates that while ROI-based masking might lead to higher reconstruction errors (because it focuses on more complex brain signals rather than easy-to-reconstruct non-brain regions), it ultimately encourages the model to learn more discriminative and functionally relevant representations.
Region-level attribution analysis further revealed that brain volumes within the limbic region and cerebellum contributed most significantly to the model’s ability to reconstruct masked areas and form effective representations. These findings align with previous ADHD research implicating these regions in attentional control. The study concludes that incorporating anatomical structure into the masking process not only enhances the interpretability of the model but also leads to more robust and discriminative representations for functional neuroimaging applications.
Also Read:
- Advancing Medical Image Analysis with Adaptable Foundation Models
- OmniBrainBench: A New Benchmark for AI in Brain Imaging Analysis
Future work will involve evaluating this approach on larger and more diverse neuroimaging datasets to assess its generalizability, as well as developing new loss functions specifically designed for region-aware reconstruction objectives. This research marks a significant step towards developing more powerful and interpretable foundation models for understanding the human brain. You can read the full research paper here.
